On Model Explanations with Transferable Neural Pathways